Two models are better than one: supervised autonomy in Swarm

Swarm is our test bed for a simple question: how far can you push autonomous AI if you stop trusting a single model to grade its own homework? The honest answer to "let one agent run free" is that it does great until it doesn't, and then it does something confidently wrong at full speed.

Split the roles

In Swarm, a master agent plans and executes against a task. A separate critique panel (independent model calls) reviews each step for mistakes and scope drift before things go too far. It's the same reason code review exists: the person who wrote the bug is the worst-positioned to see it.

Put it in a box

The whole thing runs inside a Docker sandbox, so the agent gets a real filesystem and shell to work with but can't touch the host. A bad command is contained instead of catastrophic. That single design choice is what makes it comfortable to let the agent actually do things rather than just suggest them.

What we're learning

Supervision catches more than you'd expect, and the failures it doesn't catch are usually about unclear goals, not model intelligence. The interesting frontier isn't a smarter single agent: it's better structure around ordinary ones. That's a theme in most of our AI work: keep the edges boring, keep a review path, make rollback trivial.